Supplementary material for: Cropp, A. (2026). The limits of ecological decodability of a geospatial foundation-model embedding: a registered evaluation of AlphaEarth over shallow Great Barrier Reef habitats.

Supplementary material for: Cropp, A. (2026). The limits of ecological decodability of a geospatial foundation-model embedding: a registered evaluation of AlphaEarth over shallow Great Barrier Reef habitats. EarthArXiv preprint, https://doi.org/10.31223/X50V35 (version 1 appeared under the title "What can satellite embedding fields tell us about shallow coral-reef habitats? A decodability survey of AlphaEarth over the Great Barrier Reef"; same preprint record and DOI). Contents (one zip, alphaearth-gbr-decodability-supplement-v2.zip): SUPPLEMENT.md — supplementary text §S1–§S20 (also provided typeset as SUPPLEMENT.pdf): register channels not carried in the catalogue, the Holm-adjusted table, channel and prediction correlation matrices, the Acropora construct comparison, the quiet-year maximum-power search grid, per-source label-to-channel mappings, methodological notes, a worked audit-trail example, the rubble prevalence-conditioned lead, the post-confirmatory revision rounds (A–I), the post-registration validation campaign (rounds J–N, external transfer and temporal limits), auxiliary layers, the two-axis demonstration, optical-baseline ablation detail, robustness detail, persistence-mechanism notes, ground-truth reliability and the ceiling construction, and study-design detail. A numbering note maps the v1.1 preprint's section citations onto this ordering. figS1_reliability_ceiling.png/.pdf and figS2_optical_baseline.png/.pdf — supplementary figures S1 (ground-truth reliability and the decoding ceiling) and S2 (optical-baseline ablation and Atlas-agreement circularity); these were main-text figures 5 and 6 of the v1.1 preprint. tableS1_… to tableS17_… (CSV) — all supplementary tables, generated mechanically from the study's structured run ledger; no value is transcribed by hand. tableS16 is the Holm multiplicity sweep over the post-registration campaign; tableS17 maps every campaign number quoted in the main text to its certified tableS16 row or ledger row, with a status column. run_ledger.csv — the complete machine-readable ledger of every analysis run (script, output, test, scale, metric, value, CI, p, n, datasets consumed; 3,919 rows). supp_stats.txt — computed one-off statistics quoted in §S7. The public reproduction scripts and released decoder coefficients, and the archive of the exact AlphaEarth embedding windows used, are deposited separately; their DOIs are listed as related identifiers on this record. Changes in v2 (from the published v1.1): the accompanying manuscript was restructured and retitled at preprint version 2 (same EarthArXiv record and DOI); sections §S14–§S20 and supplementary figures S1–S2 were added, carrying material condensed out of the v2 main text; supplement sections were reordered (topical notes first, review history last) with a numbering note for readers of the v1.1 preprint; tableS17 (quoted-number map) was added; five site-grain calibration ledger rows were added (ledger now 3,919 rows); a typeset SUPPLEMENT.pdf is included alongside the markdown; cross-reef counting language was clarified (the 67 pooled units are reef surveys — survey tables — not 67 distinct named reefs). This research was supported by the Commonwealth through an Australian Government Research Training Program Scholarship, undertaken through the PhD@Work arrangement at James Cook University hosted by OSIBOT PTY LTD.